3 ms·
> Ironically, the solution to this problem is more AI, not less. The phone needs to understand mirrors in order to make smarter HDR blending decisions. No, the
by TeMPOraL 3y ago
> Ironically, the solution to this problem is more AI, not less. The phone needs to understand mirrors in order to make smarter HDR blending decisions.
No, the solution is for the phone to stop blending multiple distinct exposures into one photo, because that's the crux of the problem - that's the "AI transformation" right there, decorrelating the photo from reality.
> HN has a strangely luddite tendency when it comes to photography.
It's not about luddite tendencies (leaving aside the fact that Luddites were not against technology per se - they were people fucked over by capitalist business owners, by means of being automated away from their careers). It's about recognizing that photos aren't just, or even predominantly, art. Photos are measurements, recorded observations of reality. Computational photography is screwing with that.
- foldr 3y agoNo general-purpose camera, analogue or digital, functions particularly well as a measurement tool. They are simply not designed for that purpose. Blending multiple exposures is not an "AI transformation". The first iPhone to do HDR was the iPhone 4, introduced in 2010. If you dislike HDR for some reason, no-one is forcing you to use it. There are lots of iPhone apps that will let you shoot a single exposure. That's never going to be the default in a consumer point-and-shoot camera, though, because it will give worse results for 99% of people 99% of the time.
- autoexec 3y agoCellphones have been capable of HDR for a very very long time and yet none of those older phones were capable of producing a picture like the one in the article. The problem here was not HDR. The problem was using AI to create a composite image by selectively erasing the subject of the photo from parts of the image, detecting that same person found in other photos, then cutting the subject out of those other images and pasting them into the deleted parts of the original picture before filling in the gaps to make it look like it was all one picture. This was absolutely an "AI transformation" as the original article correctly pointed out: > “It’s made like an AI decision and it stitched those two photos together,”
- foldr 3y ago>Cellphones have been capable of HDR for a very very long time and yet none of those older phones were capable of producing a picture like the one in the article. I don't think this is true? You can easily get an image like this just by choosing a different keyframe for different areas of the photo when combining the stack of exposures. No AI needed. Nor is there any 'selective erasing'. You talk about 'detecting the same person found in other photos', which suggests that you're not aware that HDR involves blending a stack of exposures (where the woman may have been in various different poses) that are part of the same 'photo' from the user's point of view (i.e. one press of the shutter button). There is no reason at all to think that the phone is actually scanning through the existing photo library to find other images of the same person. I doubt that would even be computationally feasible, given how quickly the iPhone renders the final image after the photo is taken.
- autoexec 3y agoHDR in older phones would have only merged/blended the images resulting in ghosting (the reflections would have shown the woman as having more than two arms) or blurring because they weren't using AI to select and cut out parts of the original image to swap in parts from other images. The "other photos" weren't older pictures pulled from her photo library, but were the photos taken in quick succession using different exposures. Those are still "other images of the same person" even if they are only temporarily stored in memory, but using them for exposure adjustment isn't a problem. Any old HDR capable phone would have done that. The problem is AI detecting people and deleting them to paste in "better" versions from those other photos. HDR is additive, not destructive.
- foldr 3y agoYou really don’t need any kind of fancy AI algorithm to associate different keyframes with different areas of the photo. This could be done using any number of standard computational photography techniques based on straightforward criteria such as exposure, motion blur, etc. etc. For this reason I’m skeptical of your claim that ‘HDR in older phones would have only merged/blended the images resulting in ghosting…or blurring’. Even if we assume this is true for the sake of argument, you undermine your own point here, as ghosting and blurring are also distortions of reality. You’re apparently ok with a photo showing a person with ‘more than two arms’, but not ok with a photo combining slightly different time-slices of the same scene (which is actually any photo taken using a focal plane shutter above flash sync speed :)). You're entitled to your subjective preference there, but I certainly don't share it myself.
- kaba0 3y ago> No, the solution is for the phone to stop blending multiple distinct exposures into one photo, because that's the crux of the problem - that's the "AI transformation" right there, decorrelating the photo from reality. Thanks, but I will take a proper picture over a blurry noisy mess. And please learn a bit about the whole field of measurement theory before you comment bullshit. Taking multiple measurements is literally the most basic way to get better signal to noise ratio. A blurry photo is just as much an alteration of reality, there is no real-world correspondence to my blurred face. Apple just does it more intelligently, but in rare cases messes up. It’s measurement. AI-generated stuff like a fake moon is not done by apple, and is a bad thing I also disapprove of. Smart statistics inside cameras are not at all like that.
- TeMPOraL 3y ago> Taking multiple measurements is literally the most basic way to get better signal to noise ratio. Yes. That is not a problem. Stacking those and blending them in correctly isn't a problem either - information is propagated in well-defined, and usually not surprising ways. If you're doing it, you know what you'll get. Selectively merging photos by filling different areas from different frames, and using ML to paper over the stitches, is where you start losing your signal badly. And doing that to unsuspecting people is just screwing with their perception of reality. Yes, AI generated stuff like fake Moon is even worse, but my understanding of the story in question is that the iPhone did a transformation that's about halfway between the HDR/astrophotography stuff and inpainting the Moon. And my issue isn't with the algorithms per se - hey, I'd love to have a "stitch something nice out of these few shots" button, or "share to img2img Stable Diffusion" button. My issue is when this level of AI transformation happens silently, in the background, in spaces and contexts where people still expect to be dealing with photographs, and won't notice until it's too late. And secondarily, with this becoming the default, as the default with wide-enough appeal tends to eventually become the only thing available.